Skip to content
ignisignis

Fine-tunes

A fine-tune is a catalog entry of a model family whose weights are another party’s post-training of the family’s checkpoint. It is the same family, so it loads as the family’s model does, and a client tells it apart by its served id. It is published under its own license, which may not be the family’s, and the entry states it: model list, model download, the Server view and GET /v1/models/catalog show it (the catalog). An entry of a fine-tune whose id names no family states the family it is of, family: qwen3.8-27b, which is how its 16 GB build answers to its id on a 16 GB card (ADR 0074).

The built-in catalog has two fine-tunes of the 27B: the uncensored variant, whose post-training is an abliteration, and Swift 1.5.

Swift 1.5

Section titled “Swift 1.5”

Swift 1.5 is UkisAI’s post-training of Qwen3.8-27B. It comes as two artifacts, one for each card class, and a client names the same id on both:

Card Served as Repository Size
32 GB or larger swift-1.5-27b gpillon/Swift-1.5-Qwen3.8-27B-nvfp4-ignis 18.1 GiB
16 GB swift-1.5-27b-16g gpillon/Swift-1.5-Qwen3.8-27B-trellis-3bpw-ignis 12.5 GiB 
ignis-server model download swift-1.5-27b   # prints the license, then fetches the artifact this card loads
ignis-server --model-id swift-1.5-27b       # found under --download-path, fetched when it is not there

On a 32 GB card or larger, and on a card whose class is not known, it is served as swift-1.5-27b, so a client can tell it from the 27B. On a 16 or 24 GB card, and on any RTX 30 or RTX 40 card, the same commands fetch and start swift-1.5-27b-16g, which answers to swift-1.5-27b there, as qwen3.8-27b-16g answers to qwen3.8-27b. A start, a switch and a request naming swift-1.5-27b are served by it, the answer’s model names it, and GET /v1/models lists the 32 GB entry as unavailable, the reason naming the entry that serves it. A script that must have the 16 GB artifact names swift-1.5-27b-16g itself. Both entries are pinned file by file in the generated reference.

Its license

Section titled “Its license”

Both entries are under UkisAI’s Swift Open License v1.0, swift-open-license-1.0, not Apache-2.0. Each links the LICENSE in its repository at the pinned revision; read it and the NOTICE beside it before you download.

  • Commercial use at or above the threshold is not licensed. The license’s threshold is gross revenue of US$1,000,000 or more, measured over the most recently completed fiscal year of the user together with every entity that controls it, is controlled by it, or is under common control with it. Commercial use at or above it needs a separate written license from UkisAI, the Swift Enterprise License.
  • Qualified non-profits are exempt for non-commercial or research use: the threshold does not apply to a qualified non-profit organization (a 501(c)(3) organization, or an equivalent non-profit in another jurisdiction) using the weights for non-commercial or research purposes.
  • The Qwen3.8-27B base parts remain Apache-2.0. The license limits only UkisAI’s contribution and what contains or derives from it, as these weights do.

ignis-server model download prints the license before it fetches, on a 16 GB card:

swift-1.5-27b-16g: license swift-open-license-1.0: https://huggingface.co/gpillon/Swift-1.5-Qwen3.8-27B-trellis-3bpw-ignis/blob/3fe94555a30b3df89a283bf61338281243642166/LICENSE

No step asks to accept it: whoever downloads is bound by it, as with any download from the hub (ADR 0078).

What differs from the 27B

Section titled “What differs from the 27B”
  • locate is refused, and point and box answer by the digit chain. No pointing head and no locate calibration was made on Swift’s weights, and each Swift artifact has its own weights id and so its own content hash, which the 27B’s calibrations are keyed to. On either artifact /v1/decide answers locate with 422 locate_uncalibrated, and point and box by the digit chain.
  • Vision works as on the 27B. Swift’s vision tower is the base’s, byte for byte, and each artifact loads it as the 27B’s artifact of its card class does (vision).
  • The rest is the 27B’s: the chat template and tokenizer, which Swift ships unchanged, reasoning, tool calls, the profiles and the VRAM plan. The DFlash2 drafter is the 27B’s, trained for the base and used unchanged.

How it was made

Section titled “How it was made”
  • swift-1.5-27b, weights id nvfp4full-swift-1.5: the 247 NVFP4 matrices of its text layers encoded from Swift’s BF16 checkpoint and their activation divisors calibrated locally, with the 27B NVFP4 artifact’s recipe and object list. Its DFlash2 drafter (66 objects), vision tower (333) and frontend (6) are SHA-256-equal to the 27B’s artifact’s.
  • swift-1.5-27b-16g, weights id trellis-sc30-swift-1.5: its body and output head are an exllamav3 1.5.3 conversion of Swift’s BF16 at the per-matrix bit widths of the 16 GB 27B, and every other object is the Swift NVFP4 artifact’s, byte for byte.

Each has its own content hash, so state retained under another 27B artifact is never reused for it.

The gates

Section titled “The gates”

Measured on one RTX 5090 against Swift’s own BF16: swift-1.5-27b as the 27B is served (eight lanes, a 2,048-token prefill chunk, DFlash2 with 7 draft tokens), swift-1.5-27b-16g held to a 16 GB card’s budget (rtx-16gb, 17,179,869,184 bytes, 262,144 tokens of context, two lanes, a 1,024-token chunk, vision off). The KLD limits are 1.2 times what the 27B’s artifact of the same card class scores against the base’s BF16 (0.04916 and 0.04034), registered before the run. One machine, one run per leg.

Gate swift-1.5-27b swift-1.5-27b-16g
KLD from Swift’s BF16, 16 windows of 4,096 tokens (code, prose, chat), 95% CI 0.05355 (0.04692-0.06023); limit 0.0590 0.04255 (0.03709-0.04785); limit 0.0484
Decode tok/s, DFlash2 on against off, every cell (8,192, 65,536 and 253,440 tokens of context) on above off in every cell, 3.71x to 5.27x; one lane at 8,192: 368.3 against 81.1 on above off in every cell, 3.91x to 5.22x; one lane at 8,192: 410.7 against 92.6
Smokes at temperature 0: a chat turn, a reasoning turn split into reasoning and content, a tool call pass pass

The KLD rule is the point estimate: swift-1.5-27b’s CI ends at 0.06023, above its limit; swift-1.5-27b-16g’s whole CI is under its own. The controls, against those two values: the 27B’s artifact given Swift’s BF16 scores 0.05651 and 0.04643, and each Swift artifact given the base’s BF16 0.05923 and 0.04961.

Reported, not gates:

swift-1.5-27b swift-1.5-27b-16g
DFlash2 tokens per round, recorded agent turns replayed, against the 27B’s artifact of the same class 1.001 (95% CI 0.998-1.004) 1.003 (95% CI 1.000-1.006)
DFlash2 tokens per round on its own, exact / traffic-weighted 3.467 / 3.535 (the 27B’s: 3.397 / 3.466) 3.482 / 3.552 (the 27B’s: 3.420 / 3.487)
Turns caught in a reasoning loop 11 of 32 (the 27B’s: 14 of 32, on another build and context) not measured

The replayed turns were sampled from the 27B, not from Swift. A vision gate, MMLU-Pro and a refusal comparison were not run. The builds and the gates in detail: the NVFP4 finding and the 16 GB finding.